tests: batch of small measured trims (~5–8 min slow suite, ~1–2 min fast suite)

Nobody has claimed this yet.

Small, independent trims with measured costs (per-test durations from run 29419740259, py3.12). Each can be its own small PR (or a few grouped); every PR should paste before/after `--durations` numbers.

- [ ] **`tests/test_initvals.py::test_sample_map`** — 12 parametrizations sample the full ~4000-row cavanagh dataset incl. `find_MAP`, with no assertions beyond "it ran" (195 s total). Subsample the data (~500 rows) and add a real assertion that MAP-based initvals were applied. → ~60 s.
- [ ] **`tests/test_hssm.py::test_resampling`** — 40.6 s of two real MCMC runs to assert trace-identity semantics (`_inference_obj` replaced vs reused). Assert the semantics with an attached precomputed idata + a mocked sampler instead. → seconds.
- [ ] **`tests/test_save_load.py::test_save_load_vi_mcmc`** — 28.8 s of live numpyro + double ADVI to test file round-tripping (related: #894 reports hangs here). Attach precomputed idata / tiny VI result; round-tripping doesn't need real inference. → seconds.
- [ ] **`tests/slow/test_missing_data_and_deadline_vi.py`** — uses `niter=1000` where the sibling VI files use `niter=100`, with identical isinstance-only assertions. Align to 100. → ~1.5 min.
- [ ] **`tests/slow/test_missing_data_and_deadline_mcmc.py`** — builds a *second* HSSM model per grid row solely to assert a construction-time ValueError. Hoist into one dedicated non-parametrized test. → ~40–80 s.
- [ ] **`tests/addm/test_addm_subclass.py`** — unmarked real-MCMC tests currently run in the fast PR suite (~1–2 min per fast job × 3 Pythons × every PR push). Add `@pytest.mark.slow`.
- [ ] **`tests/test_plotting.py`** — the 3 `predictive_style` parametrizations rebuild the model and re-sample an identical 2-draw posterior predictive. Share via fixture. → ~20 s.
- [ ] **aDDM cartoon/PPC/continuation tests** (`tests/addm/test_addm_cartoon.py` 64 s, `test_addm_ppc.py` 52 s, `test_addm_continuation.py` 48 s) — each rebuilds + compiles + samples the same throwaway aDDM fit. Share a module/package-scoped fixture or a canned `.nc` posterior (pattern: `tests/fixtures/idata_cavanagh_cartoon.nc`). → ~80–100 s saved.

Combined estimate: ~5–8 min per slow run, plus ~1–2 min off every fast PR job.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Start by selecting one independent item from the named test files, then run the relevant tests with pytest's `--durations` reporting to confirm the baseline. Read the existing fixtures and assertions, especially `tests/fixtures/idata_cavanagh_cartoon.nc` where relevant. Done means the selected test still verifies its stated behavior and the issue's measured before/after runtime improvement is pasted in the PR.

Written by the indexing model from the issue text.